Liam Cawley
Regression transformers matched iterative solvers on their training distribution but failed under coordinate permutations. The controls exposed dependence on a fixed ordering of feature variances and overturned the original CG identification claim.
Liam Cawley
A stability audit for fitted scikit-learn linear and logistic regression models. Computes coefficient ranges and label-flip tests over a loss-tolerance set, with explicit comparisons against sampling and quadratic approximations.
Hugh van Deventer, Itamar Pres, Liam Cawley
Sentiment-steering experiments comparing learned directions with token-unembedding baselines. The evaluator now counterbalances answer positions, and a retained Qwen recheck measures a preset intervention against an unsteered baseline.
Liam Cawley
A working note on allocating LoRA ranks under a parameter budget. The corrected derivation and reproducible matrix tests establish properties of a weighted approximation problem; the earlier unverified model-performance tables are withdrawn.
Liam Cawley
A nanoGPT notebook and deployment notes for a single Linux host with three A10 GPUs. The corrected CPU smoke configuration has been run; distributed-scaling and collective-performance results are not part of the retained record.
Sophia Tesic, Liam Cawley, Alexandra Lavacek, Namitha John, Dennis Lin
A course project exploring direct image-to-image models for 2× super resolution. Review of the retained notebook exposed training/evaluation overlap and a pixel-range error; the published record now includes corrected evaluation code and reproducible checks.
Liam Cawley
Conversation-history poisoning implementations with corrected diagnostics and matched clean-history controls. A retained three-prompt Qwen check rejected the fabricated advice; earlier simulator-generated statistics are not treated as model results.
Liam Cawley, Gabe Ronan
Hardware-in-the-loop phased-array calibration using discrete search, cached measurements, and a CNN surrogate. The tested configurations achieved a 6.5× reduction in beam-synthesis error relative to the SMI baseline.
Liam Cawley
A CIFAR-10 course project comparing CNN configurations with Cutout and Mixup and a custom ResNet with Shake-Shake. The report records configuration-level accuracy differences, without a controlled ranking of the regularizers.
Liam Cawley
A 1.6 m MFD Crosswind Mini build with a Pixhawk Cube Orange, camera payload, and autonomous takeoff and waypoint flights. The accompanying design report now separates the completed build from proposed SUAS mission capabilities.